markdown export: convert a location set instead of whole shards

[tamnd]

Closes the second ship gate box on #67.

The gap

The recovery pass in notes/Spec/2114/03-acquisition-and-crawl.md ends with a set of index locations for the pages we are missing, and nothing turned that set into a Markdown Parquet dataset. fetch --batch streams the records but only as text or JSON, and convert has neither --lang nor --dedup-digest. The only way to finish the pass was markdown export --shards over every shard those pages happen to live in. For a few thousand pages scattered across a few hundred WARC files that is on the order of a thousand times the bytes the pages are worth.

What this adds

markdown export --locations <file|-> reads the location JSONL that columnar locations and index emit, and converts exactly those records with coalesced ranged GETs.

ccrawl columnar locations --crawl CC-MAIN-2026-30 --lang vie --status 200 -o jsonl \
  | ccrawl markdown export --locations - --lang vie --dedup-digest --push=false --out ./md

Everything after the fetch is the pipeline a shard export already uses: the same extractor, the same language filter, the same digest dedup, the same schema, the same ledger and resume. That is deliberate, since a recovery pass that produced a slightly different corpus from an export of the same pages would not be a recovery of anything. The pack now takes a recordSource rather than a reader, and the WARC path and the location path share every stage after that.

New flags on markdown export:

Flag	Default	Does
--locations	(off)	Convert the records in this JSONL location stream, - for stdin
--part-size	50000	Locations per parquet part
--gap	1 MiB	Coalesce records closer together than this into one ranged read
--max-span	16 MiB	Cap on one coalesced ranged read

A part is to a location run what a shard is to a full export: the unit that gets one parquet file, one ledger entry, and one digest dedup set. The stream is cut in order, so the same input cuts the same way every time and a ledger from an interrupted run still means what it said.

A location that will not fetch is skipped rather than failing the part. A recovery pass runs against an index that can disagree with the archive, and a run that dies on the first disagreement never finishes.

--locations bypasses --shards, --source-kind, and the manifest fetch, and it is a usage error with the wet extractor, since WET files have no record offsets to point at.

warc_bytes on a location run is what the ranged reads actually pulled off the wire, holes between coalesced records included, rather than a shard size. That is the number worth publishing here, because comparing it against the shards those records live in is the whole argument for the flag.

Tests

• TestPackLocationsConvertsOnlyWhatWasAskedFor: six of twelve records in a served WARC, asserting the filler records never appear, that the six coalesce into fewer requests than one per record, and that warc_bytes lands between zero and the whole file.
• TestPackLocationsAppliesLangAndDedup: the same location set with and without --dedup-digest and --lang, checking the drop count and that the kept rows are all the requested language.
• TestPackLocationsSurvivesABadLocation: one location pointing at a file that is not there, and the part still comes back with the records that were.
• TestReadLocationParts and TestReadLocationPartsErrors: the cut into parts including the short tail, order preservation, and the four ways the flag is used wrongly.

gofmt, go vet, go test -race ./..., golangci-lint run, and scripts/docs-drift.sh are all clean. Documented in reference/markdown.md and reference/cli.md.